Benchmarking of a commercial Linux cluster

Cluster computing has emerged as a viable alternative to massively parallel supercomputers to cope with the computing demands of scientists and HPC application developers. Cluster machines, built on top of Commodity Off The Shelf components and based on open source software are becoming very popular. It is thus important to characterize the performance of these machines and to investigate the scalability of numerical algorithms when large numbers of processors are allocated. In this paper, we present a performance characterization of a large Linux cluster, that is, the IBM NetFinity located at the Maui High Performance Computing Center. The aim of this study is to evaluate the performance benefits to a few numerical applications resulting from specialized high bandwidth communication components, i.e., Myrinet switches, over a typical Beowulf cluster based on Ethernet networking. Our performance characterization study focused on the behavior of a few numerical kernels and of a climate model benchmark whose performance relies on complex iterations among various parallel algorithms. We investigated the scalability of the benchmarks under various hardware and software configurations. The study of overheads on the overall performance of a few numerical kernels deriving from different processor allocation policies concludes this paper.